[Speculative Decoding] Enabling bonus token in speculative decoding for KV cache based models (#5765)

This commit is contained in:
sroy745
2024-07-10 16:02:47 -07:00
committed by GitHub
parent 44cc76610d
commit ae151d73be
14 changed files with 645 additions and 80 deletions
+4 -1
View File
@@ -1,4 +1,4 @@
from typing import List, Optional, Tuple
from typing import List, Optional, Set, Tuple
import torch
@@ -20,6 +20,9 @@ class MLPSpeculatorWorker(NonLLMProposerWorkerBase, MultiStepWorker):
self,
execute_model_req: ExecuteModelRequest,
sample_len: int,
# Unused parameter. MLPSpeculatorWorker does not use the KV Cache and
# therefore does not need this parameter.
seq_ids_with_bonus_token_in_last_step: Set[int],
) -> Tuple[List[SamplerOutput], bool]:
"""Run the model forward pass to generate sample_len future tokens.
Returns the list of sampler output, one per layer, along with indicator